Towards trustworthy civil aviation hazards identification: An uncertainty-aware deep learning framework

Abstract

Accurate and trustworthy hazards identification is crucial for preventing accidents and ensuring flight safety. However, deep learning-based identification methods are limited by their black-box characteristics to provide trustworthy and interpretable results. Existing research on interpretable civil aviation hazard identification focuses on developing interpretable modules to be embedded in deep learning models to give engineering meaning to the results; or inferring the logic of the model’s decision-making based on the results. However, there is limited research on how to quantify and explain the uncertainty in the results. Quantifying and decomposing uncertainty not only provides confidence of results but also helps to identify the sources of unknown factors in the data, thereby providing guidance for improving model interpretability. Therefore, this paper proposes an uncertainty-aware deep learning framework for trustworthy civil aviation hazards identification. Firstly, a Bayesian multi-scale attention convolutional neural network with an integrated Monte Carlo dropout mechanism was designed, which can estimate the uncertainty of model predictions through internal randomness, thereby endowing the network with the uncertainty-aware ability. Secondly, a set of uncertainty quantification and decomposition schemes was established, which can achieve the confidence representation of the identification results and the separation of epistemic uncertainty and aleatoric uncertainty. Finally, an adjustable uncertainty decision threshold was constructed, which can be dynamically adjusted according to the risk level of application scenarios to achieve the optimal risk management. In out-of-distribution test scenarios with unknown hazards, comparisons with existing identification methods demonstrate that the proposed framework has superior uncertainty-aware capabilities and potential for engineering application.